What public data actually says about wallets that end in profit

Public write-ups about “winning Polymarket wallets” usually arrive as a fingerprint: hold for this many hours, copy this leaderboard, trade at 4 a.m., fade this category.

The pages we could re-open do not give that fingerprint.

They give a short list of documented facts, one documented activity in a dated sample, and a longer list of things that are UNKNOWN. None of this is a method for joining the tail.

What this page does not claim. It does not claim a hold-time, a copy-trade edge, an off-peak hour, or a preferred event type. It does not say “bots made $40 million.” It does not say 99% lose, or that Akey found 69% lost. It does not tell you how to join the 76.5%, the 840 addresses, or Table 1. It does not claim the April 2026 ladder, or the 2024–25 extraction total, still looks like this on 14 August 2026. How many addresses are in realized profit, and why 16% and 30% both appear, is a companion piece.

Profits sit in a thin slice of the people already winning

A CEPR discussion-paper abstract (Akey, Grégoire, Harvie, Martineau; published 12 June 2026) does not give a loss-rate percentage. It does say this, on 588 million trades and $67 billion in volume: the top 1% of users with positive PnL capture 76.5% of profits.

Read the denominator. That is the top 1% of users who already have positive PnL, not the top 1% of all users.

Andrey Sergeenkov’s all-address realized-PnL study (data through 1 April 2026, about 2.5 million addresses) is a different cut of the same idea. 15.9% of addresses were in realized profit. 2% cleared $1,000 lifetime. 840 addresses — 0.033% — cleared $100,000. The population rate, and why another write-up sits near 30%, is the companion piece. The point here is the shape of the tail.

The same Akey abstract says successful traders provide liquidity with limit orders that resolve favorably relative to realized outcomes, and unsuccessful traders take with market orders; that monthly performance is “modestly persistent,” which “may represent sample selection rather than skill”; and that insider trading is unlikely to explain the largest winners. The full PDF was not opened. Extra percentages attributed to this paper on blogs — including a “69% lost” line — are UNKNOWN from primary and are not used here.

Official docs (opened this run) match the vocabulary, not the finding: every user-facing order is a limit order; a “market order” is a limit priced to cross. Makers are never charged a taker fee. That is mechanism. It is not a study of who ends in profit, and it is not a market-making lesson.

Safe sentence, if you need one on air:

In the opened CEPR abstract (Akey et al., published 12 June 2026), the top 1% of users with positive PnL capture 76.5% of profits, on 588 million trades and $67 billion in volume. Successful traders make (limit); unsuccessful traders take (market). That is not a loss rate for all users, and it is not a how-to.

Maker flow is concentrated. That is not the same claim as “winners are market makers.”

Qin and Yang’s v1 tape (2022-11-21 to 28 April 2026) publishes a different concentration number. In their Table 5, the top 1% of maker addresses account for 84.1% of maker-side volume (Gini 0.9698; 1,235,881 maker addresses). The top 1% of taker addresses account for 69.7% of taker-side volume (2,612,688 taker addresses).

That is volume, not PnL. It sits next to Akey’s make-versus-take sentence without proving that the same addresses are the 76.5%. Do not collapse the two tables.

The same paper’s Table 13 is a weak information finding, not a winner recipe: first large (≥$100, p90) trades on resolved v1 markets hit at 0.5229 against a 0.5018 baseline. A directional-run cut adds essentially nothing (0.5018).

Many of the high-monthly-profit addresses do not last

Sergeenkov does not publish hold times in hours. He publishes activity duration, in months, for addresses whose average monthly realized profit clears a threshold.

Among about 6,600 addresses with average monthly profit above $5,000 (that is 0.26% of his ~2.5 million, data through 1 April 2026), he writes that 53% earned that average in a single month, 73% were active no more than two months, and 172 addresses — 2.6% of that subset — were active more than a year.

His table for that same >$5,000/month column, re-opened on the page: 1 month 3,517 (53.4%); 2 months 1,314 (20%); 3 months 593 (9%); 4–6 months 671 (10.2%); 7–12 months 316 (4.8%); 13+ months 172 (2.6%). The 13-month rung on the neighboring columns is 431 addresses (1.4%) at >$1,000/month and 104 (3.2%) at >$10,000/month.

On a different cut, still his: in any single month, 0.98% of all addresses earned more than $5,000; two months in a row 0.1%; three 0.03%; four 0.015%.

He also writes that every number is an address, not necessarily a person. One desk can look like many winners. One person who switches wallets can look like two short streaks instead of one long one.

That is duration of activity, not hold time of a position. The page does not say winners hold for 14 hours, or 27, or 72.

Complementary-price extraction is a documented 2024–25 activity. It is not a folklore bot story.

Saguillo, Ghafouri, Kiffer, and Suarez-Tangil (AFT 2025 / arXiv:2508.03474) opened in HTML on this run. The AFT PDF timed out; every figure below is from the HTML. They classify historical fills on markets that resolved 1 April 2024 to 1 April 2025. They are measuring whether complementary prices (YES + NO, or a mutually exclusive YES set) and a small set of logically dependent market pairs were traded in a way their method counts as extraction.

They report a realized estimate of $39,587,585.02 across the strategies they classify, with ε = $1 profit per trade. Their abstract rounds that to “40 million USD.” They do not subtract fees; they write that Polymarket did not charge per trade in that window. They say the legs are non-atomic, so one side can fill and the other can fail. They grouped a user’s bids inside a 950-block (~1 hour) window to call them one opportunity. That hour is a method parameter, not a hold-time finding.

The 2026 fee-schedule docs (opened this run) charge takers on most categories. Qin and Yang date that reform to January–March 2026. Saguillo’s sample ends 1 April 2025. Do not carry their total into a 2026 fee-on book.

That is a documented activity in a named sample. It is not “bots made $40 million.” It is not instructions.

The same paper publishes Table 1: the top ten accounts by their classified extraction, as truncated hex. The first row is $2,009,631.76 over 4,049 transactions. They say some of the large accounts show “bot-like behaviour in the number of bids.” They do not prove those addresses are bots. This packet does not invent biographies for the rows.

They also name @Tutaaa91, who they say bought both YES and NO for less than $0.02 each in a single trade they score at $58,983.36. That is one paper-reported fill. It is not a profile of a person.

What the opened pages do not say

They do not say how long a winning position is held. UNKNOWN.

They do not say that copying a profitable wallet, or a leaderboard, is profitable after slippage and fees. UNKNOWN. An official public leaderboard exists (PNL or volume; day / week / month / all; at most 50 rows). That is a list of examples, not a population study and not a copy-trade result.

They do not say winners trade off-peak, or between 03:00 and 06:00 UTC. UNKNOWN.

They do not say what types of events wallets that end in profit prefer. Sergeenkov has no event-type table. The Akey abstract has none. Saguillo’s Politics-versus-Sports plots describe where their classified extraction sat in that 2024–25 sample, not a general winner menu. Tsang and Yang’s 2024-election participation splits (~40% exclusive to Trump; 71.8% touched Trump YES; 19.4% in all four major Trump/Harris tokens) are one event, through 6 November 2024. As a population trait, event preference is UNKNOWN.

They do not say that 99% lose. Sergeenkov’s all-address red share is 84.1%, and that sentence belongs to the companion piece.

They do not say how to get into the 15.9%, the 840 addresses, or Table 1.

They do not say the April 2026 ladder, or the 2024–25 extraction total, still looks like this on 14 August 2026. We did not rerun the queries.

If you came here because a thread offered a winner recipe, the useful move is smaller than picking a new recipe: keep the facts that are on the page, keep the one dated activity in its sample window, and refuse the rest until a primary says it.